drivers/md/dm-rq.c

   1 /*
   2  * Copyright (C) 2016 Red Hat, Inc. All rights reserved.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm-core.h"
   8 #include "dm-rq.h"
   9
  10 #include <linux/elevator.h> /* for rq_end_sector() */
  11 #include <linux/blk-mq.h>
  12
  13 #define DM_MSG_PREFIX "core-rq"
  14
  15 /*
  16  * One of these is allocated per request.
  17  */
  18 struct dm_rq_target_io {
  19         struct mapped_device *md;
  20         struct dm_target *ti;
  21         struct request *orig, *clone;
  22         struct kthread_work work;
  23         blk_status_t error;
  24         union map_info info;
  25         struct dm_stats_aux stats_aux;
  26         unsigned long duration_jiffies;
  27         unsigned n_sectors;
  28         unsigned completed;
  29 };
  30
  31 #define DM_MQ_NR_HW_QUEUES 1
  32 #define DM_MQ_QUEUE_DEPTH 2048
  33 static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
  34 static unsigned dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
  35
  36 /*
  37  * Request-based DM's mempools' reserved IOs set by the user.
  38  */
  39 #define RESERVED_REQUEST_BASED_IOS      256
  40 static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
  41
  42 unsigned dm_get_reserved_rq_based_ios(void)
  43 {
  44         return __dm_get_module_param(&reserved_rq_based_ios,
  45                                      RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
  46 }
  47 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
  48
  49 static unsigned dm_get_blk_mq_nr_hw_queues(void)
  50 {
  51         return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
  52 }
  53
  54 static unsigned dm_get_blk_mq_queue_depth(void)
  55 {
  56         return __dm_get_module_param(&dm_mq_queue_depth,
  57                                      DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
  58 }
  59
  60 int dm_request_based(struct mapped_device *md)
  61 {
  62         return queue_is_mq(md->queue);
  63 }
  64
  65 void dm_start_queue(struct request_queue *q)
  66 {
  67         blk_mq_unquiesce_queue(q);
  68         blk_mq_kick_requeue_list(q);
  69 }
  70
  71 void dm_stop_queue(struct request_queue *q)
  72 {
  73         blk_mq_quiesce_queue(q);
  74 }
  75
  76 /*
  77  * Partial completion handling for request-based dm
  78  */
  79 static void end_clone_bio(struct bio *clone)
  80 {
  81         struct dm_rq_clone_bio_info *info =
  82                 container_of(clone, struct dm_rq_clone_bio_info, clone);
  83         struct dm_rq_target_io *tio = info->tio;
  84         unsigned int nr_bytes = info->orig->bi_iter.bi_size;
  85         blk_status_t error = clone->bi_status;
  86         bool is_last = !clone->bi_next;
  87
  88         bio_put(clone);
  89
  90         if (tio->error)
  91                 /*
  92                  * An error has already been detected on the request.
  93                  * Once error occurred, just let clone->end_io() handle
  94                  * the remainder.
  95                  */
  96                 return;
  97         else if (error) {
  98                 /*
  99                  * Don't notice the error to the upper layer yet.
 100                  * The error handling decision is made by the target driver,
 101                  * when the request is completed.
 102                  */
 103                 tio->error = error;
 104                 goto exit;
 105         }
 106
 107         /*
 108          * I/O for the bio successfully completed.
 109          * Notice the data completion to the upper layer.
 110          */
 111         tio->completed += nr_bytes;
 112
 113         /*
 114          * Update the original request.
 115          * Do not use blk_mq_end_request() here, because it may complete
 116          * the original request before the clone, and break the ordering.
 117          */
 118         if (is_last)
 119  exit:
 120                 blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
 121 }
 122
 123 static struct dm_rq_target_io *tio_from_request(struct request *rq)
 124 {
 125         return blk_mq_rq_to_pdu(rq);
 126 }
 127
 128 static void rq_end_stats(struct mapped_device *md, struct request *orig)
 129 {
 130         if (unlikely(dm_stats_used(&md->stats))) {
 131                 struct dm_rq_target_io *tio = tio_from_request(orig);
 132                 tio->duration_jiffies = jiffies - tio->duration_jiffies;
 133                 dm_stats_account_io(&md->stats, rq_data_dir(orig),
 134                                     blk_rq_pos(orig), tio->n_sectors, true,
 135                                     tio->duration_jiffies, &tio->stats_aux);
 136         }
 137 }
 138
 139 /*
 140  * Don't touch any member of the md after calling this function because
 141  * the md may be freed in dm_put() at the end of this function.
 142  * Or do dm_get() before calling this function and dm_put() later.
 143  */
 144 static void rq_completed(struct mapped_device *md)
 145 {
 146         /* nudge anyone waiting on suspend queue */
 147         if (unlikely(wq_has_sleeper(&md->wait)))
 148                 wake_up(&md->wait);
 149
 150         /*
 151          * dm_put() must be at the end of this function. See the comment above
 152          */
 153         dm_put(md);
 154 }
 155
 156 /*
 157  * Complete the clone and the original request.
 158  * Must be called without clone's queue lock held,
 159  * see end_clone_request() for more details.
 160  */
 161 static void dm_end_request(struct request *clone, blk_status_t error)
 162 {
 163         struct dm_rq_target_io *tio = clone->end_io_data;
 164         struct mapped_device *md = tio->md;
 165         struct request *rq = tio->orig;
 166
 167         blk_rq_unprep_clone(clone);
 168         tio->ti->type->release_clone_rq(clone, NULL);
 169
 170         rq_end_stats(md, rq);
 171         blk_mq_end_request(rq, error);
 172         rq_completed(md);
 173 }
 174
 175 static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
 176 {
 177         blk_mq_delay_kick_requeue_list(q, msecs);
 178 }
 179
 180 void dm_mq_kick_requeue_list(struct mapped_device *md)
 181 {
 182         __dm_mq_kick_requeue_list(dm_get_md_queue(md), 0);
 183 }
 184 EXPORT_SYMBOL(dm_mq_kick_requeue_list);
 185
 186 static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
 187 {
 188         blk_mq_requeue_request(rq, false);
 189         __dm_mq_kick_requeue_list(rq->q, msecs);
 190 }
 191
 192 static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
 193 {
 194         struct mapped_device *md = tio->md;
 195         struct request *rq = tio->orig;
 196         unsigned long delay_ms = delay_requeue ? 100 : 0;
 197
 198         rq_end_stats(md, rq);
 199         if (tio->clone) {
 200                 blk_rq_unprep_clone(tio->clone);
 201                 tio->ti->type->release_clone_rq(tio->clone, NULL);
 202         }
 203
 204         dm_mq_delay_requeue_request(rq, delay_ms);
 205         rq_completed(md);
 206 }
 207
 208 static void dm_done(struct request *clone, blk_status_t error, bool mapped)
 209 {
 210         int r = DM_ENDIO_DONE;
 211         struct dm_rq_target_io *tio = clone->end_io_data;
 212         dm_request_endio_fn rq_end_io = NULL;
 213
 214         if (tio->ti) {
 215                 rq_end_io = tio->ti->type->rq_end_io;
 216
 217                 if (mapped && rq_end_io)
 218                         r = rq_end_io(tio->ti, clone, error, &tio->info);
 219         }
 220
 221         if (unlikely(error == BLK_STS_TARGET)) {
 222                 if (req_op(clone) == REQ_OP_DISCARD &&
 223                     !clone->q->limits.max_discard_sectors)
 224                         disable_discard(tio->md);
 225                 else if (req_op(clone) == REQ_OP_WRITE_SAME &&
 226                          !clone->q->limits.max_write_same_sectors)
 227                         disable_write_same(tio->md);
 228                 else if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
 229                          !clone->q->limits.max_write_zeroes_sectors)
 230                         disable_write_zeroes(tio->md);
 231         }
 232
 233         switch (r) {
 234         case DM_ENDIO_DONE:
 235                 /* The target wants to complete the I/O */
 236                 dm_end_request(clone, error);
 237                 break;
 238         case DM_ENDIO_INCOMPLETE:
 239                 /* The target will handle the I/O */
 240                 return;
 241         case DM_ENDIO_REQUEUE:
 242                 /* The target wants to requeue the I/O */
 243                 dm_requeue_original_request(tio, false);
 244                 break;
 245         case DM_ENDIO_DELAY_REQUEUE:
 246                 /* The target wants to requeue the I/O after a delay */
 247                 dm_requeue_original_request(tio, true);
 248                 break;
 249         default:
 250                 DMWARN("unimplemented target endio return value: %d", r);
 251                 BUG();
 252         }
 253 }
 254
 255 /*
 256  * Request completion handler for request-based dm
 257  */
 258 static void dm_softirq_done(struct request *rq)
 259 {
 260         bool mapped = true;
 261         struct dm_rq_target_io *tio = tio_from_request(rq);
 262         struct request *clone = tio->clone;
 263
 264         if (!clone) {
 265                 struct mapped_device *md = tio->md;
 266
 267                 rq_end_stats(md, rq);
 268                 blk_mq_end_request(rq, tio->error);
 269                 rq_completed(md);
 270                 return;
 271         }
 272
 273         if (rq->rq_flags & RQF_FAILED)
 274                 mapped = false;
 275
 276         dm_done(clone, tio->error, mapped);
 277 }
 278
 279 /*
 280  * Complete the clone and the original request with the error status
 281  * through softirq context.
 282  */
 283 static void dm_complete_request(struct request *rq, blk_status_t error)
 284 {
 285         struct dm_rq_target_io *tio = tio_from_request(rq);
 286
 287         tio->error = error;
 288         blk_mq_complete_request(rq);
 289 }
 290
 291 /*
 292  * Complete the not-mapped clone and the original request with the error status
 293  * through softirq context.
 294  * Target's rq_end_io() function isn't called.
 295  * This may be used when the target's clone_and_map_rq() function fails.
 296  */
 297 static void dm_kill_unmapped_request(struct request *rq, blk_status_t error)
 298 {
 299         rq->rq_flags |= RQF_FAILED;
 300         dm_complete_request(rq, error);
 301 }
 302
 303 static void end_clone_request(struct request *clone, blk_status_t error)
 304 {
 305         struct dm_rq_target_io *tio = clone->end_io_data;
 306
 307         dm_complete_request(tio->orig, error);
 308 }
 309
 310 static blk_status_t dm_dispatch_clone_request(struct request *clone, struct request *rq)
 311 {
 312         blk_status_t r;
 313
 314         if (blk_queue_io_stat(clone->q))
 315                 clone->rq_flags |= RQF_IO_STAT;
 316
 317         clone->start_time_ns = ktime_get_ns();
 318         r = blk_insert_cloned_request(clone->q, clone);
 319         if (r != BLK_STS_OK && r != BLK_STS_RESOURCE && r != BLK_STS_DEV_RESOURCE)
 320                 /* must complete clone in terms of original request */
 321                 dm_complete_request(rq, r);
 322         return r;
 323 }
 324
 325 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
 326                                  void *data)
 327 {
 328         struct dm_rq_target_io *tio = data;
 329         struct dm_rq_clone_bio_info *info =
 330                 container_of(bio, struct dm_rq_clone_bio_info, clone);
 331
 332         info->orig = bio_orig;
 333         info->tio = tio;
 334         bio->bi_end_io = end_clone_bio;
 335
 336         return 0;
 337 }
 338
 339 static int setup_clone(struct request *clone, struct request *rq,
 340                        struct dm_rq_target_io *tio, gfp_t gfp_mask)
 341 {
 342         int r;
 343
 344         r = blk_rq_prep_clone(clone, rq, &tio->md->bs, gfp_mask,
 345                               dm_rq_bio_constructor, tio);
 346         if (r)
 347                 return r;
 348
 349         clone->end_io = end_clone_request;
 350         clone->end_io_data = tio;
 351
 352         tio->clone = clone;
 353
 354         return 0;
 355 }
 356
 357 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
 358                      struct mapped_device *md)
 359 {
 360         tio->md = md;
 361         tio->ti = NULL;
 362         tio->clone = NULL;
 363         tio->orig = rq;
 364         tio->error = 0;
 365         tio->completed = 0;
 366         /*
 367          * Avoid initializing info for blk-mq; it passes
 368          * target-specific data through info.ptr
 369          * (see: dm_mq_init_request)
 370          */
 371         if (!md->init_tio_pdu)
 372                 memset(&tio->info, 0, sizeof(tio->info));
 373 }
 374
 375 /*
 376  * Returns:
 377  * DM_MAPIO_*       : the request has been processed as indicated
 378  * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
 379  * < 0              : the request was completed due to failure
 380  */
 381 static int map_request(struct dm_rq_target_io *tio)
 382 {
 383         int r;
 384         struct dm_target *ti = tio->ti;
 385         struct mapped_device *md = tio->md;
 386         struct request *rq = tio->orig;
 387         struct request *clone = NULL;
 388         blk_status_t ret;
 389
 390         r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
 391         switch (r) {
 392         case DM_MAPIO_SUBMITTED:
 393                 /* The target has taken the I/O to submit by itself later */
 394                 break;
 395         case DM_MAPIO_REMAPPED:
 396                 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
 397                         /* -ENOMEM */
 398                         ti->type->release_clone_rq(clone, &tio->info);
 399                         return DM_MAPIO_REQUEUE;
 400                 }
 401
 402                 /* The target has remapped the I/O so dispatch it */
 403                 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
 404                                      blk_rq_pos(rq));
 405                 ret = dm_dispatch_clone_request(clone, rq);
 406                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
 407                         blk_rq_unprep_clone(clone);
 408                         blk_mq_cleanup_rq(clone);
 409                         tio->ti->type->release_clone_rq(clone, &tio->info);
 410                         tio->clone = NULL;
 411                         return DM_MAPIO_REQUEUE;
 412                 }
 413                 break;
 414         case DM_MAPIO_REQUEUE:
 415                 /* The target wants to requeue the I/O */
 416                 break;
 417         case DM_MAPIO_DELAY_REQUEUE:
 418                 /* The target wants to requeue the I/O after a delay */
 419                 dm_requeue_original_request(tio, true);
 420                 break;
 421         case DM_MAPIO_KILL:
 422                 /* The target wants to complete the I/O */
 423                 dm_kill_unmapped_request(rq, BLK_STS_IOERR);
 424                 break;
 425         default:
 426                 DMWARN("unimplemented target map return value: %d", r);
 427                 BUG();
 428         }
 429
 430         return r;
 431 }
 432
 433 /* DEPRECATED: previously used for request-based merge heuristic in dm_request_fn() */
 434 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
 435 {
 436         return sprintf(buf, "%u\n", 0);
 437 }
 438
 439 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
 440                                                      const char *buf, size_t count)
 441 {
 442         return count;
 443 }
 444
 445 static void dm_start_request(struct mapped_device *md, struct request *orig)
 446 {
 447         blk_mq_start_request(orig);
 448
 449         if (unlikely(dm_stats_used(&md->stats))) {
 450                 struct dm_rq_target_io *tio = tio_from_request(orig);
 451                 tio->duration_jiffies = jiffies;
 452                 tio->n_sectors = blk_rq_sectors(orig);
 453                 dm_stats_account_io(&md->stats, rq_data_dir(orig),
 454                                     blk_rq_pos(orig), tio->n_sectors, false, 0,
 455                                     &tio->stats_aux);
 456         }
 457
 458         /*
 459          * Hold the md reference here for the in-flight I/O.
 460          * We can't rely on the reference count by device opener,
 461          * because the device may be closed during the request completion
 462          * when all bios are completed.
 463          * See the comment in rq_completed() too.
 464          */
 465         dm_get(md);
 466 }
 467
 468 static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
 469                               unsigned int hctx_idx, unsigned int numa_node)
 470 {
 471         struct mapped_device *md = set->driver_data;
 472         struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
 473
 474         /*
 475          * Must initialize md member of tio, otherwise it won't
 476          * be available in dm_mq_queue_rq.
 477          */
 478         tio->md = md;
 479
 480         if (md->init_tio_pdu) {
 481                 /* target-specific per-io data is immediately after the tio */
 482                 tio->info.ptr = tio + 1;
 483         }
 484
 485         return 0;
 486 }
 487
 488 static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
 489                           const struct blk_mq_queue_data *bd)
 490 {
 491         struct request *rq = bd->rq;
 492         struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
 493         struct mapped_device *md = tio->md;
 494         struct dm_target *ti = md->immutable_target;
 495
 496         if (unlikely(!ti)) {
 497                 int srcu_idx;
 498                 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
 499
 500                 ti = dm_table_find_target(map, 0);
 501                 dm_put_live_table(md, srcu_idx);
 502         }
 503
 504         if (ti->type->busy && ti->type->busy(ti))
 505                 return BLK_STS_RESOURCE;
 506
 507         dm_start_request(md, rq);
 508
 509         /* Init tio using md established in .init_request */
 510         init_tio(tio, rq, md);
 511
 512         /*
 513          * Establish tio->ti before calling map_request().
 514          */
 515         tio->ti = ti;
 516
 517         /* Direct call is fine since .queue_rq allows allocations */
 518         if (map_request(tio) == DM_MAPIO_REQUEUE) {
 519                 /* Undo dm_start_request() before requeuing */
 520                 rq_end_stats(md, rq);
 521                 rq_completed(md);
 522                 return BLK_STS_RESOURCE;
 523         }
 524
 525         return BLK_STS_OK;
 526 }
 527
 528 static const struct blk_mq_ops dm_mq_ops = {
 529         .queue_rq = dm_mq_queue_rq,
 530         .complete = dm_softirq_done,
 531         .init_request = dm_mq_init_request,
 532 };
 533
 534 int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
 535 {
 536         struct request_queue *q;
 537         struct dm_target *immutable_tgt;
 538         int err;
 539
 540         md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
 541         if (!md->tag_set)
 542                 return -ENOMEM;
 543
 544         md->tag_set->ops = &dm_mq_ops;
 545         md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
 546         md->tag_set->numa_node = md->numa_node_id;
 547         md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE;
 548         md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
 549         md->tag_set->driver_data = md;
 550
 551         md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
 552         immutable_tgt = dm_table_get_immutable_target(t);
 553         if (immutable_tgt && immutable_tgt->per_io_data_size) {
 554                 /* any target-specific per-io data is immediately after the tio */
 555                 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
 556                 md->init_tio_pdu = true;
 557         }
 558
 559         err = blk_mq_alloc_tag_set(md->tag_set);
 560         if (err)
 561                 goto out_kfree_tag_set;
 562
 563         q = blk_mq_init_allocated_queue(md->tag_set, md->queue, true);
 564         if (IS_ERR(q)) {
 565                 err = PTR_ERR(q);
 566                 goto out_tag_set;
 567         }
 568
 569         return 0;
 570
 571 out_tag_set:
 572         blk_mq_free_tag_set(md->tag_set);
 573 out_kfree_tag_set:
 574         kfree(md->tag_set);
 575         md->tag_set = NULL;
 576
 577         return err;
 578 }
 579
 580 void dm_mq_cleanup_mapped_device(struct mapped_device *md)
 581 {
 582         if (md->tag_set) {
 583                 blk_mq_free_tag_set(md->tag_set);
 584                 kfree(md->tag_set);
 585                 md->tag_set = NULL;
 586         }
 587 }
 588
 589 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
 590 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
 591
 592 /* Unused, but preserved for userspace compatibility */
 593 static bool use_blk_mq = true;
 594 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
 595 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
 596
 597 module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
 598 MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
 599
 600 module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
 601 MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");